Skip to content

FiscalMindset/algsochvicky

BranchesTags

Repository files navigation

Vicky Kumar

Vicky Kumar

AI Engineer · Full-Stack Developer · Agentic Systems Builder

GitHub GitHub algsoch LinkedIn Portfolio

Building agentic AI systems, open-source integrations, and real-world automation tools.

TL;DR
Who AI Engineer specializing in agentic systems & on-device AI
What I Build Multi-agent pipelines, offline-capable apps, workflow automation
Stack Python · Kotlin · TypeScript · LangGraph · RunAnywhere SDK · Coral
Open Source 12 PRs to Coral MCP
Achievements Pull Shark (22+ PRs merged) · YOLO · Quickdraw (< 5 min merge)

Philosophy

"I build software systems, AI-native products, and agentic interfaces that turn ideas into usable, operational products."

  • AI should operate inside a product, not beside it. A useful system is not just model output. It is the interface, the workflow, the state model, and the decisions around trust.

  • Good AI UX is engineering work. Latency, control, explainability, failure states, response structure, and operator confidence are implementation concerns, not polish afterthoughts.

  • Workflow design matters more than prompt cleverness. The strongest systems are built around routes, actions, validation, and output quality, not one-off prompting tricks.

  • Applied intelligence should feel calm and exact. Serious products communicate precision through restraint, hierarchy, and interface clarity, not through noise.

Featured Projects

🏆 On-Device AI Learning — algsoch

algsoch — Android AI Study Companion

  • 100% Offline — AI runs entirely on-device via RunAnywhere SDK
  • 7 Learning Modes — Direct, Explain, Notes, Theory, Creative, Answer, Direction
  • SmolLM2-360M + SmolVLM-256M running locally on Android
  • Built with Kotlin + Jetpack Compose + RunAnywhere SDK
  • YouTube Demo
graph LR
    A[User Question] --> B[RunAnywhere SDK]
    B --> C[SmolLM2-360M<br/>On-Device LLM]
    C --> D[AI Response<br/>Mode-Adapted]
Loading

📰 Multi-Agent AI Newsroom — algsochnews

algsochnews — AI-Powered News Pipeline

  • 5-Agent orchestration — Article Extraction → News Editor → Visual Packaging → QA → Video Generation
  • Parallel processing with conditional retry routing
  • Broadcast-native output — Screenplay JSON, timed visuals, MP4 video
  • Built with LangGraph + FastAPI + React
  • Live: Frontend · API

Demo

flowchart LR
    A[Article URL] --> B[Extraction Agent]
    B --> C[News Editor Agent]
    C --> D[Visual Packaging]
    D --> E[QA Agent]
    E -->|Pass| F[Video Generation]
    E -->|Retry| C
    E -->|Retry| D
Loading

🏥 Healthcare AI — careops

careops — Coral-Powered Family Care Coordination

  • 9 Coral sources joined via single SQL interface
  • Generates doctor-ready visit packets from scattered medical records
  • Timeline synthesis across prescriptions, lab reports, symptoms, appointments
  • Safety guardrails — never diagnoses or prescribes
  • Built with Next.js 15 + Coral SQL + TypeScript
  • 22 tests passing (13 careops + 9 coral-cli)

Demo

flowchart TD
    subgraph App["CareOps Application"]
        UI["Next.js UI"] --> API["Next.js API Routes"]
        API --> Client["Coral CLI Client"]
        Client --> CLI["coral sql --format json"]
    end

    subgraph Coral["Coral Query Layer"]
        CLI --> JOIN["Cross-Source JOIN\n(patient_id key)"]
        JOIN --> PAT["careops_patients"]
        JOIN --> MED["careops_medications"]
        JOIN --> LAB["careops_lab_reports"]
        JOIN --> CHAT["careops_doctor_chats"]
        JOIN --> PHARM["careops_pharmacy_receipts"]
        JOIN --> SYMP["careops_symptom_logs"]
        JOIN --> APPT["careops_appointments"]
        JOIN --> OCR["careops_prescription_ocr"]
        JOIN --> NOTES["careops_family_notes"]
    end

    CLI --> Result["Joined SQL Result"]
    Result --> Agent["CareOps Packet Generator"]
    Agent --> Packet["Doctor Visit Packet"]
Loading

🔄 Workflow Automation — autopr & devalert

autopr — GitHub → Social Media Orchestration

  • Kestra workflows for event-driven automation
  • Generates platform-specific content (LinkedIn, Twitter, Instagram, WhatsApp)
  • Parallel AI generation with Gmail notifications
  • GitHub webhook-triggered, no backend server required
flowchart LR
    A[GitHub Webhook] --> B[Kestra Orchestrator]
    B --> C[AI Content Generation]
    C --> D1[LinkedIn]
    C --> D2[Twitter/X]
    C --> D3[Instagram]
    C --> D4[WhatsApp]
    C --> E[Email Notification]
Loading

devalert — LLM-Filtered Opportunity Alerts

  • Aggregates from 6 sources: GitHub, MLH, GSoC, Coral, HN, WeMakeDevs
  • LLM-powered scoring with configurable threshold
  • Telegram + Email notifications via Kestra orchestration
flowchart TD
    S1[GitHub] --> AGG[Aggregator]
    S2[MLH] --> AGG
    S3[GSoC] --> AGG
    S4[Coral] --> AGG
    S5[HN] --> AGG
    S6[WeMakeDevs] --> AGG
    AGG --> LLM[LLM Filter/Scorer]
    LLM --> T[Telegram Bot]
    LLM --> E[Email]
Loading

⏰ Developer Alerts — devalert

devalert — LLM-Filtered Opportunity Alerts

  • Aggregates from 6 sources: GitHub, MLH, GSoC, Coral, HN, WeMakeDevs
  • LLM-powered scoring with configurable threshold
  • Telegram + Email notifications via Kestra orchestration

🧠 LLM Interpretability — Synapse-Graph

Synapse-Graph — AI Autopsy Engine

  • Neural circuit discovery — trace which attention heads cause hallucinations
  • Causal ablation — O(n²) sweep to isolate defective components
  • OpenMetadata governance — tag heads as DEFECTIVE, mask at runtime
  • No retraining required — surgical fixes instead of full model updates
  • Built with PyTorch + FastAPI + Next.js
flowchart LR
    subgraph Dashboard["🎨 Operator Dashboard"]
        D["Next.js<br/>React<br/>@xyflow/react"]
    end

    subgraph Proxy["⚡ Neural Proxy (FastAPI)"]
        P["Generation + Tracing<br/>Governance + SSE<br/>HeadMaskStore"]
    end

    subgraph Generation["🔥 Generation"]
        O["Ollama<br/>(Preferred)"]
    end

    subgraph Tracing["🔍 Tracing"]
        T["HF Tracer<br/>PyTorch hooks"]
    end

    subgraph Governance["🛡️ Governance"]
        OM["OpenMetadata<br/>Topology + Lineage<br/>Tags → Masks"]
        DEF["⛔ DEFECTIVE<br/>→ Runtime Mask"]
    end

    D -->|"REST + SSE"| P
    P -->|"Generation"| O
    P -->|"Tracing"| T
    P -->|"Topology<br/>Lineage<br/>Tags"| OM
    OM -->|"tag"| DEF
Loading

Demo · Product Demo · Live Site

🎓 Voice & Command Systems

CommandBrain — RunAnywhere Command Memory

  • Offline-first command copilot that turns natural language into reusable shell commands
  • Local command memory stored in IndexedDB — favorites, macros, reminders, patterns
  • Safety classification with simulate/real execution modes
  • Built with React + TypeScript + RunAnywhere Web SDK
  • Live Demo · YouTube
flowchart LR
    A[Natural Language] --> B[RunAnywhere SDK]
    B --> C[Command Generator]
    C --> D[Safety Classifier]
    D --> E[IndexedDB Memory]
    D --> F[Execute/Simulate]
Loading

SpeakAI — Local English Practice

  • 100% on-device English speaking practice via RunAnywhere Web SDK + llama.cpp WASM
  • Browser speech immediately OR optional one-time local model download
  • Personality + practice modes with text + voice responses
  • No API keys, no server dependency
  • Live Demo

🎓 Education & Language Learning

Project Description Account
english_bot AI conversation practice with speech recognition algsoch

All Projects

📱 Mobile AI (On-Device)

Project Description Account
algsoch Android AI study companion, 7 learning modes, 100% offline FiscalMindset
algsochvicky Portfolio website deployed on Render FiscalMindset

🤖 Agentic AI Systems

Project Description Account
algsochnews Multi-agent newsroom with 5 agents + video generation FiscalMindset
careops Coral-powered family care coordination agent FiscalMindset
Synapse-Graph LLM interpretability with circuit discovery & causal ablation FiscalMindset
Cognivise Real-time adaptive tutoring with eye tracking algsoch
assistant_chatbot TDS problem solver & AI assistant, 55+ solvers algsoch

⚡ Workflow Automation

Project Description Account
autopr GitHub-to-social media via Kestra workflows FiscalMindset
devalert LLM-filtered developer opportunity alerts FiscalMindset

🎓 Education & Language Learning

Project Description Account
english_bot AI conversation practice with speech recognition algsoch
speakai On-browser English practice with RunAnywhere WASM algsoch

🔬 Machine Learning

Project Description Account
brain_tumor CNN brain tumor detection from MRI algsoch
brain_tumor_cnn Deep learning tumor classification algsoch

🌐 Web & Miscellaneous

Project Description Account
Sentinel Grid Kestra-first emergency response with dispatch orchestration · Demo FiscalMindset
Kairon NSUT smart attendance chatbot with Playwright scraping FiscalMindset
polybazar E-commerce platform algsoch
accomplish AI desktop agent for file management (contributor) algsoch
smart_terminal RunAnywhere CommandBrain — offline CLI assistant algsoch

Open Source Contributions

Coral MCP — 12 PRs Merged

Contributor to Coral — SQL-based data abstraction layer for AI agents. Integrated 8 AI providers:

Provider PR Description
Voyage AI #1115 Vector search integration
Sarvam AI #1112 Indian language TTS/STT
Cohere AI #1098 Command R integration
Mistral AI #1011 Mistral model support
OpenRouter #882 Unified API gateway
LM Studio #834 Local model serving
Ollama #798 Local LLM inference
Groq AI #754 Fast inference provider

View all PRs →

Skills & Tech Stack

AI Engineering

Category Technologies
On-Device AI RunAnywhere SDK, ONNX Runtime, WebAssembly, SmolLM2, SmolVLM
LLM & Agents LangChain, LangGraph, Prompt Engineering, RAG, Multi-Agent Systems
Model Ops LLM Evaluation, Hallucination Analysis, Response Evaluation
Vision & Speech Whisper, SmolVLM, Image Analysis, OCR, Web Speech API

Development

Category Technologies
Languages Python, Kotlin, JavaScript, TypeScript, SQL
Mobile Android, Jetpack Compose, React Native
Frontend React, Next.js, Tailwind CSS, Vite
Backend FastAPI, Node.js, PostgreSQL

Infrastructure

Category Technologies
Orchestration Kestra, GitHub Actions, Docker
Data Coral SQL, JSONL, SQLite, OpenMetadata
Deployment Render, Vercel, ngrok

Writing

Article Publication
How I Built CareOps Agent with Coral + OpenCode Medium
Cognivise — Real-Time Cognitive AI Tutor Medium

IDE & Tools
AI Coding OpenCode, Codex, AntiGravity, Kimchi, OpenClaw
Code Editor VS Code, Cursor, JetBrains IDEs
Terminal Warp, Hyper, iTerm2
Other Claude, ChatGPT, Gemini

Awards & Recognition

Achievement Details
Coral Hackathon Track 2 1st place — CareOps agent with 9 Coral sources
Pull Shark 22+ PRs merged on GitHub
YOLO Fast merge achievement
Quickdraw < 5 min merge time

Languages
Spoken English, Hindi
Programming Python, Kotlin, JavaScript, TypeScript, SQL

GitHub Stats

Account Repos Stars PRs Merged Contributions
@FiscalMindset 20 6 22 —
@algsoch 107+ 24+ 28 350+

🏆 Pull Shark (22+ PRs) · YOLO · Quickdraw (< 5 min merge)

Featured Work Highlights

Project Impact Link
🧠 CommandBrain Offline-first command memory + execution copilot, IndexedDB storage Live Demo
🎙️ SpeakAI 100% on-device English practice via RunAnywhere WASM Live Demo
📱 algsoch Android 100% offline AI, 7 learning modes, RunAnywhere SDK GitHub
📺 algsochnews 5-agent pipeline → broadcast video from any article URL Live Demo
🏥 careops 9 data sources joined via Coral SQL for family care coordination GitHub
🧠 Synapse-Graph Neural circuit discovery for LLM interpretability Live Demo
🏥 Brain Tumor Detection 97.9% accuracy CNN model on MRI scans Live Demo

Let's Connect

MIT License · Built by Vicky Kumar · Render & Vercel

About

Vicky Kumar builds intelligent products that feel engineered, not improvised.

algsochvicky.onrender.com/

Topics

ai engineer vicky algsoch

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages